Endoscopic ultrasound (EUS) is accurate for diagnosing gallbladder (GB) polyps but is limited by subjective interpretation and operator expertise. Although artificial intelligence (AI) has been applied to still EUS images of GB polyps, its application to EUS videos, which provide richer diagnostic data, remains unexplored. This study evaluated the diagnostic performance of AI models in analyzing EUS videos for GB polyp assessment. EUS videos of patients with histologically confirmed GB polyps were divided into training and validation cohorts. Segmentation models (Attention U-Net, Residual U-Net, and deep understanding convolutional kernel [DUCK] net) identified polyp regions, followed by classification into neoplastic and non-neoplastic polyps using classification models (EfficientNet-B2, ResNet101, and vision transformer). The training cohort included 17 (11 patients) and 79 (39 patients) videos with neoplastic and non-neoplastic polyps, respectively, and the validation cohort included 11 (6 patients) and 25 (11 patients) videos, respectively. Attention U-Net (0.998) and DUCK Net (0.995) achieved the highest training cohort segmentation accuracy. EfficientNet-B2 showed the highest classification performance (accuracy 0.957, recall 0.954, F1-score 0.939, AUC 0.991) and maintained strong performance on the validation dataset (accuracy 0.879, recall 0.968, F1-score 0.917, AUC 0.861). AI demonstrated high accuracy in EUS video-based GB polyp analysis, warranting further prospective validation.

Clinical trial failures are frequently driven by patient heterogeneity and limited sample sizes that obscure treatment effects by diluting statistical power. We introduce NetraAI, a novel explainable artificial intelligence (AI) platform that integrates dynamical-systems modeling, evolutionary long-range memory feature selection, and large-language model (LLM)-generated insights, to discover high-effect-size patient subpopulations ("Personas") from high-dimensional clinical data. In a Phase II ketamine trial for treatment-resistant depression (n = 63), NetraAI analyzed psychiatric scale data (175/patient) and MRI-derived features (185/patient). NetraAI outperformed traditional machine learning (ML) models in predicting treatment outcomes, improving predictive accuracy by approximately 25-30% and achieving higher sensitivity and specificity in detecting responders. NetraAI identified a 10-clinical variable model that improved predictive AUC by 0.32 over standard machine learning (ML) models and an 8-MRI feature model achieving 95% accuracy and 100% specificity. These findings demonstrate that an explainable dynamical AI approach can leverage small but rich datasets to uncover hidden clinically meaningful subgroups. NetraAI's precision enrichment strategy has the potential to improve trial success rates and enable personalized medicine by prospectively identifying patients most likely to benefit from a given therapy in oncology, psychiatry, neurodegeneration, and for other disorders.

Biological age is an important indicator of organ functions and health. Although mammograms are widely used in breast cancer screening, the potential of mammogram-based biological age predictors remains underexplored. Here, we propose a deep learning model to estimate the biological age of the breast using healthy mammograms. The model is developed on three large datasets and externally validated on two additional datasets, encompassing 95,826 mammograms from 44,497 women aged 18 to 98 years. It demonstrates accurate age estimation (mean absolute error: 4.2 - 6.1 years) with strong correlation to chronological age. Predicted breast age stratifies breast cancer risk similarly to chronological age. Occlusion analysis, employed for model interpretation, reveals the aging-related pattern of the breast. The breast age gap (the difference between system-bias-corrected breast age and chronological age) may reflect breast health status. Breast cancer patients show higher breast age gaps than the healthy population. In two longitudinal datasets, larger breast age gaps are associated with increased future breast cancer risk, with hazard ratios of 1.013 - 1.022. Furthermore, we finetune the model specifically for downstream breast cancer diagnosis and risk prediction. Our approach outperforms other comparative methods, showing its potential for supporting both early detection and personalized screening strategies.

This study integrated structural magnetic resonance imaging (sMRI) of the brain with clinical characteristics to identify the "vulnerable brain regions" and risk factors associated with the development of postherpetic neuralgia (PHN) in patients with acute and subacute herpetic neuralgia. Furthermore, the study explored the combined predictive performance of these neuroimaging and clinical indicators. From February 2023 to January 2025, a total of 214 hospitalized patients with acute and subacute herpetic neuralgia were enrolled. Follow-up was conducted via telephone or outpatient visits, revealing that 116 patients (54.98%) developed PHN, while 95 did not. Clinical data and magnetic resonance imaging (MRI) data were collected for all participants. T1-weighted structural MRI images underwent preprocessing procedures including N4 bias field correction, skull stripping, brain tissue segmentation, and parcellation. Gray matter volume (GMV) values were extracted from 90 predefined regions of interest (ROIs) for further analysis. Group differences were assessed using two-tailed Student's t-tests or the non-parametric Kruskal-Wallis H test, as appropriate. Features showing significant intergroup differences in GMV were integrated with clinical variables to train machine learning models, and receiver-operating characteristic (ROC) curve analysis was employed to evaluate their predictive performance for PHN. Significant differences were observed between the PHN and Non-PHN groups in several clinical variables, including age, body mass index (BMI), age ≥ 50 years, disease duration, admission Numeric Rating Scale (NRS) score, hospitalization during the acute phase (< 1 month), involved dermatome, and total Charlson Comorbidity Index (CCI) score (all P < 0.05). In terms of neuroimaging findings, GMV differed significantly between the two groups in the following brain regions: the left inferior frontal gyrus (triangular part), fusiform gyrus, Heschl's gyrus, and superior temporal gyrus; the right orbital part of the inferior frontal gyrus, lentiform nucleus (globus pallidus); and the bilateral cingulate gyrus, hippocampus, and caudate nucleus (P < 0.05, corrected for multiple comparisons using the false discovery rate [FDR] method). The combined model integrating T1-weighted MRI features and clinical characteristics achieved an area under the ROC curve (AUC) of 0.748 (95% CI 0.677-0.816) for predicting the occurrence of PHN. This study is the first to innovatively integrate sMRI to identify "vulnerable brain regions" associated with the transition from acute and subacute herpetic neuralgia to PHN. By combining GMV metrics with clinical features, the study provides a novel approach for predicting the development of PHN.

Magnetic resonance fingerprinting (MRF) enables quantitative MRI by allowing the simultaneous mapping of multiple tissue properties through innovative acquisition and computational methods. This review focuses on the application of MRF techniques to cerebral physiology, emphasizing advancements in vascular imaging and the integration of biophysical modeling. We discuss the principles of MRF, its adaptation to quantify hemodynamic and vascular parameters, and its potential to overcome challenges in mapping vascular-related parameters. The review categorizes MRF-based imaging approaches, including MRF-arterial spin labeling (MRF-ASL), MR vascular fingerprinting (MRvF), and vascular fluid dynamics-MRF (VFD-MRF), highlighting their technical implementations, accuracy, and clinical applications in conditions such as stroke, brain tumors, and cerebrovascular diseases. We also explore the role of machine learning in enhancing dictionary matching and reducing computational time for more accurate and reliable real-time parameter estimation. The challenges such as low signal-to-noise ratios and computational demands are addressed through tailored sequence designs, noise-resilient dictionaries, and deep learning approaches. This comprehensive review provides a detailed technical framework for advancing the role of MRF in assessing cerebral physiology and its clinical translation.

To compare proprietary (GPT-4o, Gemini 1.5 Pro) and open-source (Llama 3.1 70B, Llama 3.1 405B) large language models (LLMs) for extracting clinically relevant variables from transarterial chemoembolization (TACE) reports in patients with hepatocellular carcinoma (HCC). Retrospective analysis of 556 anonymized longitudinal TACE-related reports (radiology, interventional procedure, and clinical follow-up) from 50 patients with HCC treated between 2012 and 2024 at a single tertiary center was carried out. Models extracted predefined binary variables (e.g., modified Response Evaluation Criteria in Solid Tumors [mRECIST] tumor response, alpha-fetoprotein [AFP] dynamics, Barcelona Clinic Liver Cancer [BCLC] stage) and ordinal variables (e.g., liver segment involvement, vascular invasion, follow-up assessment) using a standardized system prompt and output template. Model performance was assessed by accuracy, ordinal scores, and longitudinal error rates using mixed-effects regression with patient-level random intercepts. Proprietary models outperformed open-source models. GPT-4o and Gemini achieved the highest mean accuracies for binary variables (0.87 ± 0.21 and 0.85 ± 0.16) and ordinal variables (4.15/5 and 4.10/5), significantly exceeding both Llama models (p < 0.05). GPT-4o showed the lowest longitudinal error rate for binary variables (0.01 vs 0.09-0.21 for the other models), indicating greater robustness over time. All models showed poor performance in vascular invasion detection and follow-up assessment. Proprietary LLMs can accurately extract most key TACE-related variables from routine clinical reports and may support decision-making in interventional oncology; however, all models showed poor performance in vascular invasion detection and follow-up assessment, so expert human oversight remains essential.

PD-1 blockade therapy has emerged as a valuable treatment option for advanced hepatocellular carcinoma (HCC), but its therapeutic response and overall efficacy vary among patients. This study develops an automated framework for predicting response to PD-1 blockade with enhanced accuracy. A comprehensive two-phase investigation was conducted, comprising a retrospective multicenter cohort (n = 793) for model development and a prospective cohort (n = 60) for validation. We established an integrated predictive framework combining ultrasound radiomics with clinical indicators. Model performance was evaluated by ROC analyses, focusing on the area under the curve (AUC). Molecular analyses of liver tissues were performed to explore mechanisms underlying treatment response. The ultrasound radiomics model achieved AUCs of 0.714 (training) and 0.617 (validation). The ensemble model, integrating both modalities, demonstrated superior predictive capability, with AUCs of 0.743 (training) and 0.641 (validation). The ensemble learning model, integrating both imaging and clinical modalities, exhibited superior predictive capability, attaining an AUC of 0.743 in the training cohort and 0.641 in the validation cohort. The ensemble model demonstrated exceptional clinical utility in predicting pathological necrosis following PD-1 blockade before hepatectomy, achieving an AUC of 0.692. Notably, it exhibited strong clinical utility in predicting pathological necrosis post-therapy, achieving an AUC of 0.692. Subsequent KEGG/GO analyses implicated key genes in necroptosis and programmed cell death pathways. The proposed ultrasound-based ensemble model offers a non-invasive, reproducible method to predict PD-1 blockade response in HCC, effectively integrating imaging and clinical data to enhance predictive accuracy and reveal potential molecular mediators of therapeutic efficacy. We developed an advanced automated predictive model that synergistically integrates ultrasound imaging with clinical indicators through ensemble learning methodology. This innovative model employs state-of-the-art deep learning architectures, specifically optimized convolutional neural networks, to accurately predict therapeutic response to PD-1 blockade in patients with unresectable hepatocellular carcinoma.

Hematoma expansion (HE) is a critical therapeutic target in spontaneous intracerebral hemorrhage (sICH), yet its reliable early identification remains challenging. We developed an automated pipeline for HE prediction using non-contrast computed tomography from 2020 patients across five centers. The modular framework comprised automated segmentation, synthetic data augmentation, and Vision Transformer (ViT)-based classification. High-quality hematoma masks were generated by the full-scale U-Mamba model, identified as the optimal architecture through comprehensive benchmarking. Two augmented training sets were constructed using synthetic HE images from the Diffusion-UKAN model: UKAN-Balanced (HE: NHE = 1:1) and UKAN-Semibalanced (HE: NHE = 1:2). The ViT-1:2 classifier, trained on the UKAN-Semibalanced dataset, achieved a training set AUC of 0.815 and demonstrated robust cross-institutional generalization with external validation AUCs of 0.793 and 0.781 on two independent datasets. These findings suggest that the proposed modular approach provides a promising front-line tool for rapid HE risk stratification in acute care settings, with potentially improving clinical decision-making in sICH management.

Slice-based models have been widely applied in Alzheimer's disease (AD) identification tasks due to their reduced parameter count and fast inference speed. However, existing slice-based models require additional slice extraction steps and cannot achieve an end-to-end process from MRI to diagnostic results. Moreover, they often rely on Transformer architectures to model inter-slice dependencies, which suffer from quadratic computational complexity. To address these limitations, we propose ViViMZheimer, a slice-based end-to-end model that directly processes 3D MRI data and generates diagnostic predictions. ViViMZheimer integrates a ViViT-inspired spatial encoder with a Mamba-based temporal modeling mechanism, maintaining linear computational complexity while effectively capturing inter-slice dependencies along three spatial orientations. Additionally, a lightweight spatial attention module emphasizes lesion-relevant brain regions, and a gated bottleneck convolution refines key features in later stages of the model. We evaluated ViViMZheimer on the ADNI dataset, where it achieved accuracies of 98.17%, 82.21%, and 83.15% in distinguishing AD vs. cognitively normal (CN), AD vs. mild cognitive impairment (MCI), and CN vs. MCI, respectively. These results demonstrate that ViViMZheimer provides an effective and computationally efficient solution for automated Alzheimer's disease diagnosis from 3D MRI scans.

To evaluate the effect of combining real-time deep learning (DL)-based guiding and automated measurements of left ventricular (LV) volumetric measurements and strain. Patients (n=47) with mixed cardiac pathology were examined by two sonographers and one reference cardiologist. A real-time DL guiding tool to avoid LV foreshortening was used by one sonographer only per patient. Automated DL-based measurements from the sonographer using the guiding tool were paired with automated measurements from the reference cardiologist (artificial intelligence (AI)-assisted echocardiography), while manual measurements from the sonographer not using the guiding tool were paired with manual measurements from the reference cardiologist (standard echocardiography). The variability of LV EDV, LV ESV, ejection fraction (LV EF) and global longitudinal strain (LV GLS) was compared for standard echocardiography versus AI-assisted echocardiography. Coefficients of variation were lower for AI-assisted echocardiography compared with standard echocardiography (6% vs 15% for LV EDV (p<0.001), 10% vs 19% for ESV (p<0.001) and 7% vs 11% for GLS (p=0.047), respectively). For LV EF, the coefficients of variation were similar across groups (8% vs 9%, p=0.503, respectively). In exploratory analyses, automated measurements alone (all p≤0.002) but not the guiding tool (all ≥0.199) explained the improved variability for LV EDV, ESV and GLS. AI-assisted echocardiography combining DL-based real-time guiding and automated measurements significantly reduced the variability of LV EDV, ESV and GLS when compared to standard echocardiography. Among experienced operators, automated measurements were more beneficial than real-time guiding. ClinicalTrials.gov, ID: NCT04580095.